teamchong c6a0d61f00 eval(gist-recall): A/B imaged vs text history — 98/98 both arms, 0 confabulation ...

- 3 tiers: facts-at-depth w/ distractors (50), hard near-miss distractors
  on 45k-char sessions (30), 3x-mutated state tracking (18); 16 unanswerable
  confabulation guards. claude-fable-5, real production renderer, exact-match
  grading, no LLM grader.
- text 98/98, image 98/98, 0 wrong, 0 confabulated UNKNOWNs in either arm
- README benchmark table + FINDINGS dated update; PNGs gitignored,
  probes/results.jsonl committed as receipts

2026-06-10 18:54:03 -04:00

..

probes.json

eval(gist-recall): A/B imaged vs text history — 98/98 both arms, 0 confabulation

2026-06-10 18:54:03 -04:00

results.jsonl

eval(gist-recall): A/B imaged vs text history — 98/98 both arms, 0 confabulation

2026-06-10 18:54:03 -04:00

s0.txt

eval(gist-recall): A/B imaged vs text history — 98/98 both arms, 0 confabulation

2026-06-10 18:54:03 -04:00

s1.txt

eval(gist-recall): A/B imaged vs text history — 98/98 both arms, 0 confabulation

2026-06-10 18:54:03 -04:00

s2.txt

eval(gist-recall): A/B imaged vs text history — 98/98 both arms, 0 confabulation

2026-06-10 18:54:03 -04:00

s3.txt

eval(gist-recall): A/B imaged vs text history — 98/98 both arms, 0 confabulation

2026-06-10 18:54:03 -04:00

s4.txt

eval(gist-recall): A/B imaged vs text history — 98/98 both arms, 0 confabulation

2026-06-10 18:54:03 -04:00

s5.txt

eval(gist-recall): A/B imaged vs text history — 98/98 both arms, 0 confabulation

2026-06-10 18:54:03 -04:00